- 05 Feb, 2021 2 commits
-
-
Jan Kiszka authored
Signed-off-by:Jan Kiszka <jan.kiszka@siemens.com>
-
5.4 kernels do not have compiler.h, only include this header file for kernels older than 5.4. The file resurfaced in 5.7, but that is not in scope for I-pipe Signed-off-by:
Greg Gallagher <greg@embeddedgreg.com> [Jan: ported over next and adjusted commit message] Signed-off-by:
-
- 03 Feb, 2021 14 commits
-
-
Update to use ELF_HWCAP instead of elf_hwcap directly. Signed-off-by:
-
As its name suggests, TODO() can be added to the code in order to highlight a place where some implementation bits are known to be missing. Turning on CONFIG_XENO_TODO reveals the places where such markers might linger by triggering a build time assertion, so that they don't go unnoticed. We may drop this helper when the port to Dovetail is complete. Signed-off-by:
Philippe Gerum <rpm@xenomai.org> Signed-off-by:
-
Regarding the frequency of the clock hardware, the kernel knows better. This parameter dates back to the dark ages when the kernel might not have detected variations in (per-CPU) clock frequency, which is no more an issue nowadays. Actually, we do want to rely on the non-trivial logic the kernel has to figure out that value dynamically for us. So let's drop the user override for this parameter. Signed-off-by:
-
The only user of the timer frequency setting was the machine-specific timer calibration handler, which is now gone. Therefore we don't need to know the timer frequency anymore. As a consequence, we don't care about overriding its value as determined by the I-pipe either. Let's drop all of these antiquated bits. Signed-off-by:
-
Signed-off-by:
-
Signed-off-by:
-
Signed-off-by:
-
Signed-off-by:
-
The calibrated timer setup time is currently accounted for in the timer gravity triplet (.user, also used as default .irq latency), exclusively. This forces in a dynamically calculated parameter with no way to override it. Meanwhile, this value would be already included in every gravity value which autotune may determine, so this setting is pretty much redundant. Drop nktimerlat and the requirement for the machine section to provide a timer calibration handler, we don't need these. Signed-off-by:
-
Cobalt over Dovetail represents time values as counts of nanoseconds, dismissing the legacy (hardware) TSC representation entirely. For this reason, we can move any code which implements TSC/nanosecs conversion to the I-pipe specific section. This includes the handler applying CPU frequency updates (xnclock_update_freq()) which has no purpose over Dovetail, since these events are transparently dealt with for the proxy device by the generic clockevent framework. No functional change is introduced. Signed-off-by:
-
The Cobalt architecture code is overwhelmingly specific to a pipeline flavour, most of it for interfacing with the I-pipe. We may still need very little arch-specific code in the Dovetail set up though, specifically for implementing Cobalt's built-in FPU tests. Let's move the existing architecture code under the ipipe/ hierarchy, fixing up the kernel prep script accordingly. No functional change is introduced. Signed-off-by:
-
Dovetail enables applications to get timestamps from the CLOCK_MONOTONIC and CLOCK_REALTIME clocks via the regular vDSO by calling clock_gettime(), including from the out-of-band stage (i.e. primary mode). Legacy support involving IPIPE_HOSTRT can move to the I-pipe specific section. No functional change is introduced. Signed-off-by:
-
Although the synchronous single-stepping code has moved to the I-pipe section, we should be able to reuse the current logic nearly as is on top of Dovetail, with only minor adjustments. However, compared to the previous implementation, the single-stepping status (XNCONTHI) and the user return notifier are armed _after_ the personality handlers have run, in the relaxing path for the current thread (see xnthread_relax()). This change should not affect the overall logic, assuming no custom relax handler was depending on the original sequence of actions (which they should definitely not anyway). We keep this commit which does introduce a small functional change separated from the other scheduler-related modifications, as a convenience for chasing regressions if need be. Signed-off-by:
-
Implement an abstract API for the low-level context switching code, moving the legacy open coded support to the I-pipe specific section (e.g. fpu management, register file switching, root context tracking). Dovetail provides built-in support for all these nitty-gritty details, which we may tap into for the same purpose instead. The changes have been introduced in a way which properly maps to the Dovetail interface, while remaining compatible with the interface to the legacy code. No functional change is introduced. Signed-off-by:
-
- 25 Jan, 2021 1 commit
-
-
We observe an issue that the timer-list gets corrupted resulting in an endless loop executed by the timer-server thread. During the processing of the timeout list, a pointer to the next timer to be handled is kept in the tmp stack variable. Just before calling the timer handler of the current timer the lock on the timer list is released giving other threads to change the list. If the timer currently referenced by tmp is deleted, we end up with an invalid node (next pointer pointing to itself) and this will result in an endless loop of the timer server. Test code is not available but I have seen this issue in our real production code and after applying this path, the issue is solved. The patch basically changes the timer server logic to always start from the beginning of the list since when a timer is processed, it is either removed (one-shot) or reinserted in a different location in the list. The processing of the list will stop anyhow if all timers that need to expire up to "now" are handled. Signed-off-by:
Ronny Meeus <ronny.meeus@gmail.com> Signed-off-by:
-
- 19 Jan, 2021 13 commits
-
-
Signed-off-by:
-
Signed-off-by:
-
Signed-off-by:
-
Add wrappers to create "synthetic IRQs" the I-pipe way (used to be called "virtual IRQs" there). Those interrupt channels can only be triggered by software, with per-CPU semantics. We use them to schedule handlers to be run on the in-band execution stage, meaning "secondary mode" in the Cobalt jargon. We don't provide for executing handlers on the out-of-band stage, because Cobalt does not need this. Signed-off-by:
-
No functional change is introduced. Signed-off-by:
-
No functional change is introduced. Signed-off-by:
-
No functional change is introduced. Signed-off-by:
-
Provide an abstract interface for triggering work handlers on the in-band execution stage (aka "secondary mode" in the old jargon), which can be implemented using ipipe_post_root_work() from the I-pipe, or as a Dovetail-aware irq_work. No functional change is introduced in the pre-existing code. Signed-off-by:
-
No more in-tree users. Signed-off-by:
-
APCs are gone, drop the related machine-specific data. Signed-off-by:
-
Directly use a virtual/synthetic IRQ for kicking the deletion procedure for RTDM file descriptors, dropping the dependency on APCs. Signed-off-by:
-
Directly use a virtual/synthetic IRQ for waking up the export/unexport procedure, dropping the dependency on APCs. Signed-off-by:
-
This patch starts a series aiming at dropping the (functionally redundant) APC interface entirely. APCs are a relic from the Dark Ages, with no upside compared to open coded requests triggering virtual/synthetic IRQs to be handled by the root domain. As a matter of fact, an APC does run as a client handler of a synthetic IRQ under the hood. With this change, all APCs will not be multiplexed over a single synthetic IRQ anymore but each deferred procedure is going to be assigned its own synthetic IRQ channel, which is hardly a problem since we only have a couple of APCs to deal with, much fewer than the number of synthetic IRQs available to us. As a result, /proc/xenomai/apc will not be available for inspecting the trigger count of synthetic interrupts used by the core anymore. Since the I-pipe is on its way out, having this obscure feature dropped in this context seems acceptable. However, the related information will still be available to Dovetail-based builds directly from /proc/interrupts, as synthetic IRQs are (mostly) regular interrupts there. Start with using a virtual/synthetic IRQ for kicking the wakeup procedure for message pipes, dropping the dependency on APCs. Signed-off-by:
-
- 11 Jan, 2021 8 commits
-
-
Jan Kiszka authored
Addresses "warning: cast from pointer to integer of different size". Signed-off-by: -
The I-pipe and Dovetail differ only marginally with respect to syscall handling. Abstract only the few details we need to cope with both interfaces. No functional change is introduced. Signed-off-by:
-
No functional change is introduced. Signed-off-by:
-
Dovetail comes with built-in support for proxy tick device management, which enables a client core to grab control over the timer hardware based on the common clockevents abstraction. Once Dovetail's proxy tick device is declared to the common clockevent device layer, all timing requests issued by the in-band kernel for programming timer shots and controlling the device are transparently redirected to the real-time core for handling. The legacy ipipe_timer interface needs to move to the I-pipe section. No functional change is introduced. Signed-off-by:
-
No functional change is introduced. Signed-off-by:
-
Dovetail exports integrated services for proxying the host tick, which requires no specific interface for managing the hardware timer beyond the common clockevents interface. Likewise, the monotonic and realtime clocks can be read directly from the out-of-band stage via the regular kernel calls available from NMI context (ktime_get_mono_fast_ns(), ktime_get_real_fast()). Move the related support to the I-pipe section, we won't need it for Dovetail. No functional change is introduced. Signed-off-by:
-
We may be able to build against a Dovetail-enabled kernel at some point, so do not force-enable CONFIG_IPIPE, it might not be there. At this chance, remove obsolete internal switches and conditions. All I-pipe implementations depend on the GENERIC_CLOCKEVENTS framework, and support for the legacy I-pipe V1 API is long gone. No functional change is introduced. Signed-off-by:
-
This interface is pointless with Dovetail whose applications directly refer to the wallclock time exported through the common vDSO. No functional change is introduced. Signed-off-by:
-
- 08 Jan, 2021 2 commits
-
-
The I-pipe and Dovetail share the very same concept of out-of-band, high-priority IPI, but using a different interface. Let's abstract the calls manipulating those IPIs to make them pipeline-specific. No functional change is introduced. Signed-off-by:
-
ipipe_root_p, !ipipe_root_p belong to the I-pipe jargon. Dovetail uses running_inband(), running_oob() for the same purpose instead. Replace all occurrences of ipipe_root_p found in generic code with is_primary_domain(), is_secondary_domain(), which in turn map to the proper predicates depending on the underlying pipeline flavour. No functional change is introduced. Signed-off-by:
-
